AI at the IMO: Gold in 2025, Reported 42/42 in 2026

Quick answer: The International Mathematical Olympiad (IMO) is the world’s top math competition for high school students: six proof problems over two days, worth 42 points. In 2024 Google DeepMind’s AI reached silver-medal level. In 2025, Gemini Deep Think and an OpenAI model each scored 35/42, gold-medal level. At IMO 2026 in Shanghai, Huawei reported that its Celia system scored 42/42 and RedNote’s team reported full marks for dots-note-3.0 under committee-organized grading, as AFP relayed.

For decades, the International Mathematical Olympiad was a benchmark that computers could not touch. Its problems need creative proofs, not calculations. Then, in just three years, AI went from silver-level scores to reported perfect scores. Here is what happened, and how to read the headlines.

What Is the IMO?

The IMO has run since 1959. Each country sends up to six students under 20. Over two days, contestants get 4.5 hours each day for three problems, in algebra, combinatorics, geometry and number theory. Each problem is worth 7 points, for a maximum of 42. Roughly the top twelfth of contestants earn gold.

Timeline: AI at the IMO

YearAI resultHow it was checked
2024Google DeepMind’s AlphaProof and AlphaGeometry 2: 28/42, silver levelFour problems fully solved. Humans translated problems into formal inputs; AlphaProof worked in Lean, AlphaGeometry 2 in its own geometry system; some solutions took up to three days. Timothy Gowers and Joseph Myers marked them.
2025Gemini Deep Think: 35/42, gold levelOfficially graded by IMO coordinators, within the time limit, in natural language
2025OpenAI experimental model: 35/42, gold levelGraded by three former IMO medalists, not by the IMO
2026Huawei Celia and RedNote dots-note-3.0: 42/42Company announcements, relayed by AFP; dots reported committee-organized grading and used Python code tools. We found no IMO-hosted report confirming both.
2026Several frontier models: 42/42 on independent runsUnofficial harnesses, some graded by AI agents

IMO 2026, the 67th, was held in Shanghai, with exam days on July 15 and 16, 2026. Among 666 human contestants, seven scored a perfect 42.

How to Read an “AI Gold” Headline

  • Who graded it? Official IMO grading is stronger evidence than self-grading or AI graders.
  • Same conditions? Humans get 9 hours total with pen, paper and drawing instruments, but no calculators or computers. Some AI runs used code tools, repeated attempts or huge computing power; equal clock time is not equal resources.
  • Could it have seen the answers? Running after the problems are released is necessary but not sufficient: check the model version, training cutoff, run date and whether it had access to published solutions.
  • Natural language or formal proof? Lean proofs can be checked by computer; written proofs need human graders.

In its July 19, 2025 statement, the IMO distinguished checking that proofs are correct from validating a company’s methods, computing power, human involvement and reproducibility. That caution still applies to every claim.

Does This Mean AI Can Do Research Math?

Olympiad problems are hard but known to have neat solutions. Research problems may have no solution at all, or need new theory. Still, AI models are now contributing research results, such as an internal OpenAI model’s disproof of Erdős’s unit distance conjecture (which overturned a predicted growth rate rather than finding the exact answer), so the gap is narrowing. For students, the IMO remains what it always was: a test of how you think.

Common Misconceptions

“AI won a gold medal.” AI systems are not official contestants and do not receive medals; reports describe gold-medal-level scores.

“Perfect scores mean the IMO is solved forever.” They mean current problems are within reach of the best systems. The competition is still for students.

Watch: AI and the Math Olympiad in Video

If you learn better by watching, these videos are worth your time.

The last IMO problem AI could not solve (3Blue1Brown and PolyaMath). A beautiful explanation of an olympiad problem and what it reveals about AI.
LLM wins Gold on 2025 Math Competition (IMO) (Caleb Writes Code). A short summary of the 2025 gold-level results.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. What is the maximum score at the IMO?
Found a bug? Let us know!
2. What level did AI reach at IMO 2024?
Found a bug? Let us know!
3. Which 2025 AI result was graded by IMO coordinators?
Found a bug? Let us know!
4. Why does official grading matter?
Found a bug? Let us know!
5. Where was IMO 2026 held?
Found a bug? Let us know!

Frequently Asked Questions

Did AI get a perfect score at the IMO?

At IMO 2026, Huawei reported 42/42 for Celia and RedNote’s team reported full marks for dots-note-3.0 under committee-organized grading, as AFP relayed. Other models reported 42/42 on unofficial runs. We found no IMO-hosted report independently confirming both.

Did AI win a gold medal at the IMO?

AI systems are not official contestants, but in 2025 Google DeepMind’s Gemini Deep Think scored 35/42, gold-medal level, under official grading. An OpenAI model scored the same, graded by former medalists.

How is the IMO scored?

Contestants solve six proof problems over two days, 4.5 hours per day. Each problem is worth 7 points, for a maximum of 42.

Can AI do research mathematics?

Increasingly, yes. In 2026, AI systems contributed to research results such as the disproof of Erdős’s unit distance conjecture, though many claims still need expert checking.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "AI at the IMO: Gold in 2025, Reported 42/42 in 2026 - Effortless Math"?

No one replied yet.

Leave a Reply